Power distribution network defect sample generation and intelligent identification method and system

Abstract

The invention discloses a power distribution network defect sample generation and intelligent identification method, and the method comprises the steps: carrying out the segmentation and extraction of foreground, background and target information of an existing sample through employing a target feature segmentation method; carrying out secondary fusion on the extracted information to synthesize a new sample; the method comprises the following steps: performing training classification on an original sample by using an internal feature cycle acquisition method to form a base-model, and only storing original data features; and training the synthesized expansion sample by using a base-model, expanding model parameter characteristics through the synthesized expansion sample, performing network structure expansion on the small sample through secondary generation of network characteristics, and finally completing generation of the defect sample of the power distribution network and construction of the intelligent identification model. According to the method, a multi-layer model fusion recognition method is constructed, and the problems that a small-scale sample construction model is low in accuracy and poor in robustness are solved.

Description

Technical Field

[0001] The present invention relates to the field of distribution network defect identification, and in particular to a distribution network defect sample generation and intelligent identification method and system. Background Art

[0002] With the application of artificial intelligence models in power distribution inspection, the analysis efficiency of power distribution inspection images has been greatly improved. However, the defects of distribution network equipment are complex and the frequency of occurrence of different defects varies. In power distribution inspection, the application of artificial intelligence models has indeed greatly improved the efficiency of image analysis, but at the same time, it also faces the problem of insufficient sample size, especially when identifying defects in distribution network equipment. The following are some key defects and challenges.

[0003] The contradiction between insufficient sample size and model requirements: The construction of an algorithm model requires a large number of defect sample libraries, but the defect samples that are the focus of attention in distribution line inspections are very small. There is a contradiction between the small sample size and the demand for large quantities of samples in model construction, making it difficult to build an algorithm model.

[0004] Limitations of traditional sample expansion methods: Current methods for sample generation and expansion mainly include traditional expansion methods such as sample rotation, translation, and blurring. Traditional expansion methods are linear processing of existing samples and cannot generate new sample features. In the process of model construction, it is easy to cause overfitting of model parameters. The detection and recognition model trained by expanding samples using traditional methods has poor robustness to changing environments.

[0005] Problems with adversarial sample generation methods: To solve the problem of building a few-sample defect intelligent identification model, the invention with application number 202110311489.4 proposes a deep neural network black-box adversarial sample generation system based on channel information. The model structure can be identified based on the power consumption trajectory using an appropriate machine learning algorithm, and then the identified model structure is used to train an equivalent model with the same decision boundary as the target model. Finally, the equivalent model is used to generate adversarial samples to attack the target model. The black-box sample generation method based on the generative adversarial network can generate new sample features based on the existing extremely small sample size, but the features generated by the adversarial network sample generation method cannot maintain consistency in the same dimension as the original sample features, there are too many noise points, and even erroneous features appear. When the adversarial network is used to generate samples for model training, the model will have problems of underfitting and failure to converge.

[0006] Challenges of defect detection system: In actual distribution network equipment defect detection, due to the complexity and diversity of defect samples and the insufficient sample size, building a high-precision and generalizable intelligent perception system faces huge challenges. Especially in key scenarios such as power grid fault diagnosis and equipment health evaluation, the problems of uneven sample distribution and scarce key samples are particularly prominent.

[0007] Limitations of intelligent inspection robots: Although intelligent inspection robots play an increasingly important role in power inspection, they still need to be improved in terms of intelligence, pattern recognition algorithms, and comprehensive diagnostic analysis capabilities. For example, some robots have deficiencies in instrument recognition capabilities, infrared temperature measurement accuracy, and resistance to strong electromagnetic interference. Summary of the invention

[0008] The purpose of the present invention is to provide a distribution network defect sample generation and intelligent identification method and system, which can expand the existing samples, complete the construction of distribution network defect sample generation and intelligent identification model, and retain the characteristics of the samples, solving the problem of low accuracy in small-scale sample construction.

[0009] In order to achieve the above object, the present invention is implemented through the following technical solutions: A method for generating and intelligently identifying distribution network defect samples comprises the following steps: Use the target feature segmentation method to segment and extract existing samples; The extracted information is fused again to synthesize new expanded samples; Design a prototype network for small sample defect recognition with multi-layer model feature fusion; Design a small sample optimization network with added feature generation structure; Through secondary generation of network features, the network structure of the small sample optimization network is expanded to complete the generation of distribution network defect samples and the construction of intelligent identification model.

[0010] Furthermore, the segmentation and extraction of samples includes segmenting and extracting the foreground, background and target information of existing samples, realizing the segmentation of the sample foreground and background through graph processing methods, and realizing the separation and extraction of the target area and the background through high-degree-of-freedom sparse matrix decomposition and maximum value estimation methods.

[0011] Furthermore, the image processing method for segmenting the sample foreground and background includes the following steps: Divide the original image into There are non-overlapping image blocks of image blocks, , Represents the number of non-overlapping image blocks. Image blocks Perform matrix vectorization operation on the RGB pixel values ​​in to form a matrix gradient map; The foreground and background are determined according to the gradient change of the matrix gradient map. If there is a gradient mutation, there is a conversion between the foreground and the background, thereby achieving the segmentation of the foreground and the background.

[0012] Furthermore, the global gradient average is first calculated; if there is a gradient value greater than the global gradient average, it is considered that there is a gradient mutation; if the mutation direction is positive, it is a mutation from background to foreground; if the mutation direction is negative, it is a mutation from foreground to background.

[0013] Furthermore, information fusion first randomly selects any two or more elements from the foreground, background and target of the extracted image, directly and irregularly splices multiple elements to form a spliced ​​image; uses the edge blurring method to fuse features on the edge side to generate new data sample features.

[0014] Furthermore, the prototype network for small sample defect recognition is a dual-channel network, which constructs different paired samples by combination, determines whether two samples belong to the same category through the distance of the top-level paired samples, and generates corresponding probabilities.

[0015] Further, As the mean center of the i-th category, N represents the number of category samples, Represents the jth sample of the i-th category. After embedding the test sample x, the distance calculation is performed , determine the category of the sample, where is the distance between sample x and the mean center of the i-th category; In the process of sample feature collection, sample collection is performed through the internal feature cycle feature network. The sample collection formula is as follows: Where t is the time step, h(t) represents the state at the tth time step, is the weight function, U is the transfer matrix, represents the sample at the tth time step, Indicates a non-zero compensation parameter.

[0016] Furthermore, the generation of the small sample optimization network includes the following steps: The original sample training classification forms the base-model, and only the original data features are stored; Use the base-model to train the synthetic augmented samples and build a small sample migration network model; Expand the model parameter features through synthetic augmented samples; A feature generation network structure is added to the small sample defect recognition prototype network with a multi-layer feature fusion structure to perform secondary expansion and generation of input image features and enrich the model feature parameters again.

[0017] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the above method when executing the computer program. A computer storage medium stores instructions, and when the instructions are executed on a computer, the computer executes the above method.

[0018] The advantages of the present invention are: Design a sample segmentation synthesis and expansion method, use the target feature segmentation method to segment and extract the foreground, background and target information of the existing samples, perform secondary fusion on the extracted information, synthesize new samples, and enrich the small sample data sample library; Design a prototype network for small sample defect recognition with multi-layer feature fusion, train and classify the original samples using the internal feature loop collection method to form a base-model, which only stores the original data features to achieve overall fitting of the small sample data set; A small sample optimization network with added feature generation structure is designed. Through secondary generation of network features, the network structure of small samples is expanded, the sample feature parameters are enriched, and the model robustness is increased. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 Schematic diagram of matrix gradient change in Example 1 of the present invention; Figure 2 This is a schematic diagram of sample category determination in Example 2 of the present invention; Figure 3 This is a schematic diagram of image segmentation and extraction in Example 3 of the present invention; Figure 4 This is a schematic diagram of replacing the edge joint value with a blurred pixel value in Example 3 of the present invention. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0021] Example 1 A method for generating and intelligently identifying distribution network defect samples comprises the following steps: Design a sample segmentation synthesis and expansion method, use the target feature segmentation method to segment and extract the foreground, background and target information of the existing samples; The extracted information is fused again to synthesize new samples; Design a small sample defect recognition prototype network with multi-layer model feature fusion, train and classify the original samples using the internal feature loop collection method to form a base-model, which only stores the original data features; Use the base-Model to train the synthesized expanded samples and build a small sample migration network model. Through the synthesized expanded samples, expand the model parameter features, design a small sample optimization network with added feature generation structure, and expand the network structure of the small sample through secondary generation of network features. Finally, complete the generation of distribution network defect samples and the construction of intelligent identification model.

[0022] Specifically, the sample segmentation and extraction includes segmenting and extracting the foreground, background and target information of the existing samples, realizing the segmentation of the sample foreground and background through graph processing methods, and realizing the separation and extraction of the target area and the background through high-degree-of-freedom sparse matrix decomposition and maximum value estimation methods.

[0023] Sample segmentation can be achieved by using graph processing methods to segment the sample foreground and background, and by using high-degree-of-freedom sparse matrix decomposition and maximum estimation methods to separate and extract the target area from the background.

[0024] Furthermore, the image processing method for segmenting the sample foreground and background includes the following steps: Please refer to Figure 1 , the original image is divided into There are non-overlapping image blocks of image blocks, , represents the number of non-overlapping image blocks, is a natural number, Image blocks The RGB pixel values ​​in the matrix are vectorized to form a matrix gradient map. The foreground and background are judged according to the gradient change of the matrix gradient map. If there is a gradient mutation, there is a conversion between the foreground and the background, thereby realizing the segmentation of the foreground and the background.

[0025] The change of the matrix gradient map is mainly manifested as a sudden drop or a gentle change of the gradient. The gradient part 1 of the gradient map belongs to a gentle change, and the gradient from gradient part 1 to gradient part 2 belongs to a sudden drop of the gradient. The original image represented by gradient part 1 and the original image represented by gradient part 2 are different backgrounds. Similarly, gradient part 3 and gradient part 4 belong to different parts.

[0026] The global gradient average is calculated and used as the gradient discrimination condition. If the gradient value is greater than the gradient average and the mutation direction is positive, it is a mutation from background to foreground; if the mutation direction is negative, it is a mutation from foreground to background.

[0027] After sample segmentation, the foreground, background and target of the image are extracted. Different foregrounds, backgrounds and targets are merged through element splicing and edge fusion methods. First, any two or more elements in the extracted image foreground, background and target are randomly selected, and multiple elements are directly spliced ​​irregularly to form a spliced ​​image. The problem that the spliced ​​image is not smooth enough at the edge intersection of different elements requires the use of edge blurring methods to fuse features on the edge side to generate new data sample features.

[0028] The prototype network for small sample defect recognition with multi-layer feature fusion is a dual-channel network. It constructs different paired samples by combination, judges whether two samples belong to the same category by the distance of the top layer paired samples, and generates the corresponding probability. As the mean center of N categories, after embedding the test sample x, the distance between it and the N centers is calculated to determine the category of the sample. In the process of sample feature collection, sample collection is performed through the internal feature cycle feature network. The sample collection formula is as follows: Where t is the time step, h(t) represents the state at the tth time step, is the weight function, U is the transfer matrix, represents the sample at the tth time step, Represents a non-zero compensation parameter, which is a hyperparameter setting, usually 1 or -1.

[0029] After obtaining the fused image and constructing a small sample defect recognition network model with multi-layer feature fusion, a feature generation network structure is added to the multi-layer feature fusion structure to perform secondary expansion and generation of the input image features, enrich the model feature parameters again, and finally complete the distribution network defect sample generation and intelligent recognition model construction.

[0030] Example 2 This example selects three categories to calculate the mean center. Please refer to Figure 2 The samples of the three categories are C ij , i=1,2,3;j∈(0,N), N is the number of sample categories, then the center of the i-th category is: Embed the test sample x and calculate the distance with the three centers.

[0031] Example 3 This embodiment takes the broken conductor in a small sample image of a power distribution network as an example to generate a small sample defect sample for power distribution and construct an identification model.

[0032] The main form of wire breakage defect is that a single wire in the distribution line is completely or partially broken. It is a critical defect in the distribution network. The breakage of the distribution line wire will directly lead to power outage of the line. In order to ensure the normal operation of the distribution network, the frequency of wire breakage must be strictly controlled, and the defect must be repaired as soon as it is discovered. Therefore, the number of defect samples is small, and it is impossible to use traditional large-sample training methods such as deep learning for model construction.

[0033] Based on the above reasons, it is necessary to expand the small sample size and build the recognition model training for the wire strand break defect. The specific steps are as follows: (1) Expand the small sample of broken wire strands and use the target feature segmentation method to segment and extract the existing samples.

[0034] The segmentation and extraction of broken wire samples includes segmenting and extracting the foreground, background and target information of existing wire samples, realizing the segmentation of sample foreground and background through graph processing methods, and realizing the separation and extraction of target area and other background through high-degree-of-freedom sparse matrix decomposition and maximum value estimation methods.

[0035] Please refer to Figure 3 The red part is the broken wire strands in the key focus area, the brown area is the sample background, the blue and yellow areas are the sample foreground, and the green part intersects with the red part, which is the sample foreground.

[0036] First, the brown area background is segmented from the target and foreground. The overall gradient of the distribution network inspection image background is relatively gentle, and there is a sudden gradient reduction feature between the foreground and the background. Based on these two properties, the background and foreground are segmented.

[0037] Obtain the gradient distribution map of the entire image: I(x,y) represents the value of the image pixel, x is the horizontal coordinate of the image, and y is the vertical coordinate of the image Then the first-order gradient is: , is the gradient of the horizontal and vertical coordinates of the image.

[0038] After gradient calculation, the gradient distribution of the image can be obtained. If there is a sudden change in the gradient between the foreground and the background, the foreground and the background can be segmented according to the suddenly changed coordinate values.

[0039] Secondly, after removing the background, it is necessary to remove the target information of broken wire strands and the sample foreground through high-degree-of-freedom sparse matrix decomposition and maximum estimation method.

[0040] The target area of ​​wire strand breakage has the characteristics of morphological continuity and gradient balance. By establishing a high degree of freedom matrix of target information and sample foreground and performing consistent decomposition on the matrix, the segmentation matrix of target information and sample foreground is obtained as follows: The part of the matrix that is all 0 is the target area, the part that is 1 is the sample foreground, and the part where 1 and 0 intersect is the overlapping part of the two.

[0041] After all backgrounds are removed, the target area is extracted as a separate area, and its target area matrix is: (2) The extracted information is fused again to synthesize new expanded samples.

[0042] Firstly, any two or more elements among the foreground, background and target of the extracted image of the broken strand of the ground conductor are randomly selected, and the multiple elements are directly and irregularly spliced ​​to form a spliced ​​image.

[0043] To solve the problem that the edges of different elements in the stitched image are not smooth enough, it is necessary to use the edge blurring method to fuse the features on the edge side and generate new data sample features.

[0044] The edge blurring method mainly calculates the average value of the eight pixel values ​​around the target stitching image to obtain the blurred pixel value, and uses the blurred pixel value to replace the value of the edge stitching, such as Figure 4 shown.

[0045] (3) Design a small sample defect recognition prototype network with multi-layer model feature fusion, expand the network structure of the small sample optimization network through secondary network feature generation, and complete the generation of distribution network defect samples and the construction of intelligent recognition model.

[0046] The prototype network for identifying small sample defects of broken ground wires with multi-layer feature fusion is a dual-channel network. Different paired samples are constructed by combination. The distance between the top paired samples is used to determine whether two samples belong to the same category and generate corresponding probabilities. c1, c2, and c3 are used as the mean centers of the three categories. After embedding the test sample x, the distance between the three centers is calculated to determine the category of the sample. Please refer to Figure 2 In the process of sample feature collection, samples are collected through the internal feature loop feature network. The sample collection formula is as follows: Where t is the time step, h(t) represents the state at the tth time step, is the weight function, U is the transfer matrix, represents the sample at the tth time step, Represents a non-zero compensation parameter, which is a hyperparameter setting, usually 1 or -1.

[0047] After obtaining the fused image and constructing a small sample defect recognition network model with multi-layer feature fusion, a feature generation network structure is added to the multi-layer feature fusion structure to perform secondary expansion and generation of the input image features, enrich the model feature parameters again, and finally complete the distribution network defect sample generation and intelligent recognition model construction.

[0048] Example 4 This embodiment provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented: Step 1: Use the target feature segmentation method to segment and extract existing samples: Divide the original image into There are non-overlapping image blocks of image blocks, , represents the number of non-overlapping image blocks, is a natural number, Image blocks The RGB pixel values ​​in the matrix are vectorized to form a matrix gradient map. The foreground and background are judged according to the gradient change of the matrix gradient map. If there is a gradient mutation, there is a conversion between the foreground and the background, thereby realizing the segmentation of the foreground and the background.

[0049] Step 2: Fuse the extracted information for the second time to synthesize new expanded samples: After sample segmentation, the foreground, background and target of the image are extracted. Different foregrounds, backgrounds and targets are merged through element splicing and edge fusion methods. First, any two or more elements in the extracted image foreground, background and target are randomly selected, and multiple elements are directly spliced ​​irregularly to form a spliced ​​image. The problem that the spliced ​​image is not smooth enough at the edge intersection of different elements requires the use of edge blurring methods to fuse features on the edge side to generate new data sample features.

[0050] Step 3: Design a small sample defect recognition prototype network with multi-layer model feature fusion: The prototype network for small sample defect recognition with multi-layer feature fusion is a dual-channel network. It constructs different paired samples by combining them, determines whether two samples belong to the same category by the distance between the top layer sample pairs, and generates the corresponding probability. As the mean center of N categories, after embedding the test sample x, the distance between it and the N centers is calculated to determine the category of the sample. In the process of sample feature collection, sample collection is performed through the internal feature cycle feature network. The sample collection formula is as follows: Where t is the time step, h(t) represents the state at the tth time step, is the weight function, U is the transfer matrix, represents the sample at the tth time step, Represents a non-zero compensation parameter, which is a hyperparameter setting, usually 1 or -1.

[0051] Step 4: Design a small sample optimization network with added feature generation structure: The fused image is obtained in the second step, and the small sample defect recognition network model of multi-layer feature fusion is constructed in the third step. Then, a feature generation network structure is added to the multi-layer feature fusion structure to perform secondary expansion and generation of the input image features, thereby enriching the model feature parameters again.

[0052] Step 5: Through secondary generation of network features, the network structure of the small sample optimization network is expanded to complete the generation of distribution network defect samples and the construction of intelligent identification model.

[0053] Example 5 This embodiment provides a computer storage medium, in which instructions are stored. When the instructions are executed on a computer, the computer executes the method steps described in Embodiment 4: Use the target feature segmentation method to segment and extract existing samples; The extracted information is fused again to synthesize new expanded samples; Design a prototype network for small sample defect recognition with multi-layer model feature fusion; Design a small sample optimization network with added feature generation structure; Through secondary generation of network features, the network structure of the small sample optimization network is expanded to complete the generation of distribution network defect samples and the construction of intelligent identification model.

[0054] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for generating and intelligently identifying distribution network defect samples, characterized in that: Includes steps: Use the target feature segmentation method to segment and extract existing samples; The extracted information is fused again to synthesize new expanded samples; Design a prototype network for small sample defect recognition with multi-layer model feature fusion; Design a small sample optimization network with added feature generation structure; Through secondary generation of network features, the network structure of the small sample optimization network is expanded to complete the generation of distribution network defect samples and the construction of intelligent identification model.

2. According to claim 1, the method for generating and intelligently identifying distribution network defect samples is characterized in that: The segmentation and extraction of samples includes segmentation and extraction of foreground, background and target information of existing samples, segmentation of sample foreground and background through graph processing methods, and separation and extraction of target area and background through high-degree-of-freedom sparse matrix decomposition and maximum estimation methods.

3. The method for generating and intelligently identifying distribution network defect samples according to claim 2 is characterized in that: The image processing method for segmenting the sample foreground and background comprises the steps of: Divide the original image into There are non-overlapping image blocks of image blocks, , Represents the number of non-overlapping image blocks. Image blocks Perform matrix vectorization operation on the RGB pixel values ​​in to form a matrix gradient map; The foreground and background are determined according to the gradient change of the matrix gradient map. If there is a gradient mutation, there is a conversion between the foreground and the background, thereby achieving the segmentation of the foreground and the background.

4. The method for generating and intelligently identifying distribution network defect samples according to claim 3 is characterized in that: First, the global gradient average is calculated; if there is a gradient value greater than the global gradient average, it is considered that there is a gradient mutation; if the mutation direction is positive, it is a mutation from background to foreground; if the mutation direction is negative, it is a mutation from foreground to background.

5. The method for generating and intelligently identifying distribution network defect samples according to claim 1, characterized in that: Information fusion first randomly selects any two or more elements from the foreground, background and target of the extracted image, directly and irregularly splices multiple elements to form a spliced ​​image; uses the edge blurring method to fuse features on the edge side to generate new data sample features.

6. The method for generating and intelligently identifying distribution network defect samples according to claim 1, characterized in that: The small sample defect recognition prototype network is a dual-channel network, which constructs different paired samples by combination, judges whether two samples belong to the same category by the distance of the top layer paired samples, and generates corresponding probabilities.

7. The method for generating and intelligently identifying distribution network defect samples according to claim 6 is characterized in that: Will As the mean center of the ith category, Represents the jth sample of the i-th category. After embedding the test sample x, the distance calculation is performed , determine the category of the sample, where is the distance between sample x and the mean center of the i-th category; In the process of sample feature collection, sample collection is performed through the internal feature cycle feature network. The sample collection formula is as follows: Where t is the time step, h(t) represents the state at the tth time step, is the weight function, U is the transfer matrix, represents the sample at the tth time step, Indicates a non-zero compensation parameter.

8. The method for generating and intelligently identifying distribution network defect samples according to claim 1, characterized in that: The generation of the small sample optimization network includes the steps of: The original sample training classification forms the base-model, and only the original data features are stored; Use the base-model to train the synthetic augmented samples and build a small sample migration network model; Expand the model parameter features through synthetic augmented samples; A feature generation network structure is added to the small sample defect recognition prototype network with a multi-layer feature fusion structure to perform secondary expansion and generation of input image features and enrich the model feature parameters again.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the method according to any one of claims 1 to 8 is implemented.

10. A computer storage medium, characterized in that: The computer storage medium stores instructions, and when the instructions are executed on a computer, the computer is caused to perform the method according to any one of claims 1 to 8.

Images